← Back to Overview
Always-On Intelligence · LPNPU · Intel 06E

88 TOPS.
Under One and a Half Watts.

The LPNPU is Eventide's intelligence layer that never sleeps. Running on the same LP Island as the LPE cores and 2DKR tile, this dedicated neural processor delivers 88 TOPS of inference power while consuming less than 1.4W at peak—enabling always-on keyword detection, presence sensing, and hardware thread orchestration without touching the main CPU.

Intel 06E
88 TOPS
Always-On
Wake-on-Voice
SONAR
88
TOPS Peak
<1.4W
Peak Power
344
TOPS Combined
<0.3W
Idle

Three Foundational Roles

01

Always-On Inference

Continuous, zero-CPU wake-on-voice keyword detection using Gaussian Mixture Model (GMM) phoneme analysis. The LPNPU monitors ambient audio in real-time, detects user intent through keyword confidence thresholds, and routes SONAR presence signals—all without involving the main processor or waking the LPE cluster.

02

Hardware Thread Director Co-Pilot

Predictive task dispatch co-processor running lightweight AI models to forecast task runqueue states before threads become schedulable. The LPNPU signals the UHP Solar Eclipse cores to pre-stage frequency targets, delivering sub-100ns arbitration advantage over OS scheduler latency.

03

System Intelligence Layer

Persistent behavioral AI that learns user habits and system patterns. Predicts app pre-staging needs, recommends dynamic power profiles, and orchestrates tile wake sequences. Runs continuously, shaping the system's response before user demand is visible to the OS.

Wake-on-Voice Execution Stack

Audio Input Ambient audio stream from always-on microphone array
LPNPU Analysis Phoneme extraction & Gaussian Mixture Model keyword detection
Threshold Cross Confidence score exceeds wake trigger threshold
Darkmont Wakeup LPNPU triggers LP E-core cluster; system responsive before user finishes speaking

WiFi SONAR: Presence Sensing

Doppler-Shift Presence Detection

SONAR (Signal Observation and Nearfield Analysis Radar) leverages the LPNPU's always-on compute to analyze WiFi signal reflections in real-time. By measuring Doppler shifts in reflected RF energy, the LPNPU detects human presence within a ~3m radius—without requiring new hardware, external cameras, or privacy intrusions.

When a user approaches the device, SONAR wakes the display from black screen at sub-100ms latency. As the user moves away, the display dims intelligently. The system learns spatial occupancy patterns, enabling predictive wake sequencing and context-aware power state transitions.

This tight coupling between WiFi analog front-end and LPNPU compute creates a new class of ambient awareness—proximity sensing without the power cost of dedicated proximity hardware.

SONAR Specifications
Detection Range ~3 meters
Wake Latency <100ms
Frequency Band WiFi 6 / 6E
Detection Method Doppler Analysis
Power Draw Negligible
Privacy Impact Zero imaging

Power Advantage vs. CPU Inference

Inference Platform Compute Location Typical Power Draw Advantage
LPNPU Dedicated NPU (LP Island) <1.4W Baseline
LPE Cores Low-Power Execution (Darkmont) ~8W 5.7× more power
DP Cores Data Path (UHP Solar Eclipse) ~22W 15.7× more power

Primary Use Cases

🎙️
Wake-on-Voice
📡
SONAR Presence
⚡
Predictive App Dispatch
🔇
Background Noise Suppression